def __init__(self, config):
super(ResNet18, self).__init__()
self.classes = config['num_classes']
self.in_shape = config['input_shape']
self.channels = 3
self.pad1 = ZeroPad2d(padding=3)
self.conv1 = Conv2d(self.channels, 64, kernel_size=(7, 7), stride=2)
self.bn1 = BatchNorm2d(64)
self.pad2 = ZeroPad2d(padding=1)
self.pool1 = MaxPool2d(kernel_size=3, stride=2)
# stage 2
self.conv_block_s2 = ResNetConv(64, filters=64, stride=1)
self.identity_s2_1 = ResNetIdentity(64, filters=64)
self.identity_s2_2 = ResNetIdentity(64, filters=64)
# stage 3
self.conv_block_s3 = ResNetConv(64, filters=128)
self.identity_s3_1 = ResNetIdentity(128, filters=128)
self.identity_s3_2 = ResNetIdentity(128, filters=128)
# stage 4
self.conv_block_s4 = ResNetConv(128, filters=256)
self.identity_s4_1 = ResNetIdentity(256, filters=256)
self.identity_s4_2 = ResNetIdentity(256, filters=256)
# stage 5
self.conv_block_s5 = ResNetConv(256, filters=512)
self.identity_s5_1 = ResNetIdentity(512, filters=512)
self.identity_s5_2 = ResNetIdentity(512, filters=512)
# final stage
self.linear = Linear(512, self.classes)
def __init__(self, base_filters=16):
super(DiscriminatorROI, self).__init__()
self.conv_layers = Sequential(
ZeroPad2d((1, 2, 1, 2)),
Conv2d(3, base_filters, kernel_size=4, stride=2, bias=False),
LeakyReLU(0.2, inplace=True), ZeroPad2d((1, 2, 1, 2)),
Conv2d(base_filters,
2 * base_filters,
kernel_size=4,
stride=2,
bias=False), BatchNorm2d(2 * base_filters, momentum=0.8),
LeakyReLU(0.2, inplace=True), ZeroPad2d((1, 2, 1, 2)),
Conv2d(2 * base_filters,
4 * base_filters,
kernel_size=4,
stride=2,
bias=False), BatchNorm2d(4 * base_filters, momentum=0.8),
LeakyReLU(0.2, inplace=True))
self.roi_pool = RoIAlign(output_size=(3, 3),
spatial_scale=0.125,
sampling_ratio=-1)
self.classifier = Sequential(
Conv2d(4 * base_filters, 1, kernel_size=3, padding=0, bias=False))
def __init__(self):
super(UpConv_7, self).__init__()
self.act_fn = nn.LeakyReLU(0.1, inplace=False)
self.offset = 7 # because of 0 padding
from torch.nn import ZeroPad2d
self.pad = ZeroPad2d(self.offset)
m = [
nn.Conv2d(3, 16, 3, 1, 0),
self.act_fn,
nn.Conv2d(16, 32, 3, 1, 0),
self.act_fn,
nn.Conv2d(32, 64, 3, 1, 0),
self.act_fn,
nn.Conv2d(64, 128, 3, 1, 0),
self.act_fn,
nn.Conv2d(128, 128, 3, 1, 0),
self.act_fn,
nn.Conv2d(128, 256, 3, 1, 0),
self.act_fn,
# in_channels, out_channels, kernel_size, stride=1, padding=0, output_padding=
nn.ConvTranspose2d(256,
3,
kernel_size=4,
stride=2,
padding=3,
bias=False)
]
self.Sequential = nn.Sequential(*m)
def __init__(self,
shape,
list_zernike_ft,
list_zernike_direct,
padding_coeff = 0.,
deformation = 'single',
features = None):
# Here we define the type of Model we want to be using, the number of polynoms and if we want to implement a deformation.
super(Aberration, self).__init__()
#Check whether the model is given the lists of zernike polynoms to use or simply the total number to use
if type(list_zernike_direct) not in [list, np.ndarray]:
list_zernike_direct = range(0,list_zernike_direct)
if type(list_zernike_ft) not in [list, np.ndarray]:
list_zernike_ft = range(0,list_zernike_ft)
self.nxy = shape
# padding layer, to have a good FFT resolution
# (requires to crop after IFFT)
padding = int(padding_coeff*self.nxy)
self.pad = ZeroPad2d(padding)
# scaling x, y
if deformation == 'single':
self.deformation = ComplexDeformation()
elif deformation == 'scaling':
self.deformation = ComplexScaling()
else:
self.deformation = Identity()
self.zernike_ft = Sequential(*(ComplexZernike(j=j + 1) for j in list_zernike_ft))
self.zernike_direct = Sequential(*(ComplexZernike(j=j + 1) for j in list_zernike_direct))
def __init__(self, N_side_in, **kwargs):
self.N_side = N_side_in
self.canv_shape = (self.N_side, self.N_side)
self.op_dict = {
'union' : union,
'rect' : self.primitive_rect
}
#'subtract' : subtract,
self.op_str_list = list(self.op_dict.keys())
#print(self.op_str_list)
self.N_ops = len(self.op_str_list)
self.N_non_primitive_ops = 1
self.N_params = 4
self.zero_pad = ZeroPad2d(1)
#self.peaky_noise = Beta(0.03*torch.ones(self.canv_shape), 0.47*torch.ones(self.canv_shape))
#self.peaky_noise = Beta(1*torch.ones(self.canv_shape), 8*torch.ones(self.canv_shape))
self.peaky_noise = Beta(0.05*torch.ones(self.canv_shape), 0.45*torch.ones(self.canv_shape))
self.canv_dist = kwargs.get('canv_dist', 'bernoulli')
assert self.canv_dist in ['bernoulli', 'beta'], 'Canv dist must be either bernoulli or beta!'
noise_methods = {
'bernoulli' : 'bern',
'beta' : 'peaky_blur',
}
self.noise_method = noise_methods[self.canv_dist]
def __init__(self, width, height, num_encoders, safety_margin=0):
self.height = height
self.width = width
self.num_encoders = num_encoders
self.width_crop_size = optimal_crop_size(self.width, num_encoders,
safety_margin)
self.height_crop_size = optimal_crop_size(self.height, num_encoders,
safety_margin)
self.padding_top = ceil(0.5 * (self.height_crop_size - self.height))
self.padding_bottom = floor(0.5 *
(self.height_crop_size - self.height))
self.padding_left = ceil(0.5 * (self.width_crop_size - self.width))
self.padding_right = floor(0.5 * (self.width_crop_size - self.width))
self.pad = ZeroPad2d((self.padding_left, self.padding_right,
self.padding_top, self.padding_bottom))
self.cx = floor(self.width_crop_size / 2)
self.cy = floor(self.height_crop_size / 2)
self.ix0 = self.cx - floor(self.width / 2)
self.ix1 = self.cx + ceil(self.width / 2)
self.iy0 = self.cy - floor(self.height / 2)
self.iy1 = self.cy + ceil(self.height / 2)
def forward(self, phonemes, spectrograms, len_phonemes, training=False):
"""
:param phonemes: (batch, alphabet, time), padded phonemes
:param spectrograms: (batch, freq, time), padded spectrograms
:param len_phonemes: list of phoneme lengths
:return: decoded_spectrograms, attention_weights
"""
spectrs = ZeroPad2d(
(0, 0, 1, 0))(spectrograms)[:, :-1, :] # move this to encoder?
keys, values = self.txt_encoder(phonemes)
queries = self.audio_encoder(spectrs)
att_mask = mask(shape=(len(keys), queries.shape[1], keys.shape[1]),
lengths=len_phonemes,
dim=-1).to(self.device)
if hp.positional_encoding:
keys += positional_encoding(keys.shape[-1], keys.shape[1],
w=hp.w).to(self.device)
queries += positional_encoding(queries.shape[-1],
queries.shape[1],
w=1).to(self.device)
attention, weights = self.attention(queries,
keys,
values,
mask=att_mask)
decoded = self.audio_decoder(attention + queries)
return decoded, weights
def __init__(self, file_shape: tuple, output_filters=8, output_channels=2):
super(Generator, self).__init__()
# DownSampling
self.down1 = UNetDown(file_shape[0], output_filters, normalize=False)
self.down2 = UNetDown(output_filters, output_filters * 2)
self.down3 = UNetDown(output_filters * 2, output_filters * 4)
self.down4 = UNetDown(output_filters * 4, output_filters * 8)
self.down5 = UNetDown(output_filters * 8, output_filters * 8)
self.down6 = UNetDown(output_filters * 8, output_filters * 8)
self.down7 = UNetDown(output_filters * 8, output_filters * 8)
self.down8 = UNetDown(output_filters * 8, output_filters * 8)
# UpSampling
self.up1 = UNetUp(output_filters * 8, output_filters * 8)
self.up2 = UNetUp(output_filters * 16, output_filters * 8)
self.up3 = UNetUp(output_filters * 16, output_filters * 8)
self.up4 = UNetUp(output_filters * 16, output_filters * 8)
self.up5 = UNetUp(output_filters * 16, output_filters * 4)
self.up6 = UNetUp(output_filters * 8, output_filters * 2)
self.up7 = UNetUp(output_filters * 4, output_filters)
self.last = nn.Sequential(
Upsample(scale_factor=(2, 4)),
ZeroPad2d((0, 0, 1, 0)),
Conv2d(output_filters * 2,
output_channels,
kernel_size=(4, 4),
stride=1,
padding=(1, 0)),
Sigmoid(),
)
def forward(self, data):
x_name, x_behavior = data.split([1000, 4000], 1)
x_name = self.embedder1(x_name)
x_behavior = self.embedder2(x_behavior)
x_name = x_name.unsqueeze(1)
x_behavior = x_behavior.unsqueeze(1)
pad = ZeroPad2d(padding=(0, 0, 2, 1))
x_name_pad = pad(x_name)
x_name_cnn1 = F.relu(self.cnn1_1(x_name)).squeeze(-1).permute(0, 2, 1)
x_name_cnn2 = F.relu(self.cnn1_2(x_name_pad)).squeeze(-1).permute(
0, 2, 1)
x_name_cnn3 = F.relu(self.cnn1_3(x_name)).squeeze(-1).permute(0, 2, 1)
x_behavior = F.relu(self.cnn2(x_behavior)).squeeze(-1).permute(0, 2, 1)
x = torch.cat([x_name_cnn1, x_name_cnn2, x_name_cnn3, x_behavior],
dim=-1)
x, (h_n, c_n) = self.lstm(x)
x = h_n[-1, :, :]
x = F.relu(self.lin1(x))
x = F.dropout(x, p=0.2, training=self.training)
x = F.relu(self.lin2(x))
x = F.dropout(x, p=0.2, training=self.training)
x = F.sigmoid(self.lin3(x))
return x
def __init__(self, in_ch, out_ch):
super(Upward, self).__init__()
self.depool = Sequential(
UpsamplingNearest2d(scale_factor=2), ZeroPad2d((1, 2, 1, 2)),
Conv2d(in_ch, out_ch, kernel_size=4, stride=1), ReLU(inplace=True),
InstanceNorm2d(out_ch))
def end_block(self):
return Sequential(
Conv2d(in_channels=256, out_channels=256, kernel_size=(1, 8), groups=256),
ZeroPad2d((4, 3, 0, 0)),
Conv2d(in_channels=256, out_channels=512, kernel_size=(1, 1)),
self.activation(),
Gate(),
Dropout(p=0.4)
)
def __init__(self, in_ch, out_ch, stride=2, normalise=True):
super(ConvBlock, self).__init__()
self.pool = Sequential(
ZeroPad2d((1, 2, 1, 2)),
Conv2d(in_ch, out_ch, kernel_size=4, stride=stride),
LeakyReLU(negative_slope=0.2, inplace=True))
if normalise:
self.pool.add_module('instance_norm', InstanceNorm2d(out_ch))
def __init__(self, in_ch, out_ch, normalise=True):
super(Downward, self).__init__()
self.pool = Sequential(ZeroPad2d((1, 2, 1, 2)),
Conv2d(in_ch, out_ch, kernel_size=4, stride=2),
LeakyReLU(negative_slope=0.2, inplace=True))
if normalise:
self.pool.add_module('batch_norm', BatchNorm2d(out_ch,
momentum=0.8))
def __init__(self, base_filters=64):
super(PatchGAN, self).__init__()
self.down1 = ConvBlock(1, base_filters, normalise=False)
self.down2 = ConvBlock(base_filters, 2 * base_filters)
self.down3 = ConvBlock(2 * base_filters, 4 * base_filters)
self.down4 = ConvBlock(4 * base_filters, 8 * base_filters)
self.validity = Sequential(
ZeroPad2d((1, 2, 1, 2)),
Conv2d(8 * base_filters, 1, kernel_size=4, stride=1))
def __init__(self, nb_classes, nb_channels, base_filters=16):
super(PatchGAN, self).__init__()
self.down1 = Downward(nb_classes + nb_channels,
base_filters,
normalise=False)
self.down2 = Downward(base_filters, 2 * base_filters)
self.down3 = Downward(2 * base_filters, 4 * base_filters)
self.padding = ZeroPad2d((1, 2, 1, 2))
self.validity = Sequential(
Conv2d(4 * base_filters, 1, kernel_size=4, stride=1), Sigmoid())
def get_transform():
transform = nn.Sequential(
ZeroPad2d(150),
RandomAffine(degrees=(-20, 20),
translate=(0.25, 0.25),
scale=(1.1, 1.5)),
)
def transform_fn(image, batch_size):
b_image = image.repeat(batch_size, 1, 1, 1)
return transform(b_image)
return transform_fn
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.data = self.data.unsqueeze(1).float().div(255)
from torch.nn import ZeroPad2d
pad = ZeroPad2d(2)
self.data = torch.stack([pad(sample.data) for sample in self.data])
self.targets = self.targets.long()
self.data = self.data.sub_(self.data.mean()).div_(self.data.std())
# self.data = self.data.sub_(0.1307).div_(0.3081)
# Put both data and targets on GPU in advance
self.data, self.targets = self.data, self.targets
print('MNIST data shape {}, targets shape {}'.format(self.data.shape, self.targets.shape))
def __init__(self, file_shape: tuple, output_filters=8):
super(Discriminator, self).__init__()
self.model = Sequential(
*self.build_block(file_shape[0] * 2,
output_filters,
normalization=False),
*self.build_block(output_filters, output_filters * 2),
*self.build_block(output_filters * 2, output_filters * 4),
*self.build_block(output_filters * 4, output_filters * 8),
ZeroPad2d((0, 0, 1, 0)),
Conv2d(output_filters * 8,
1,
kernel_size=(4, 1),
padding=(1, 0),
stride=1),
)
def __init__(self, nb_classes, nb_channels, base_filters=32):
super(Generator, self).__init__()
self.down1 = Downward(nb_classes, base_filters, normalise=False)
self.down2 = Downward(base_filters, 2 * base_filters)
self.down3 = Downward(2 * base_filters, 4 * base_filters)
self.down4 = Downward(4 * base_filters, 8 * base_filters)
self.up1 = Upward(8 * base_filters, 4 * base_filters)
self.up2 = Upward(8 * base_filters, 2 * base_filters)
self.up3 = Upward(4 * base_filters, base_filters)
self.out_conv = Sequential(
UpsamplingNearest2d(scale_factor=2), ZeroPad2d((1, 1, 1, 1)),
Conv2d(2 * base_filters, nb_channels, kernel_size=3, stride=1),
Tanh())
def forward(ctx, input, selector, weight, bias, stride=1):
# Pre-pad the input.
input = ZeroPad2d(weight.shape[-1] // 2)(input)
# Build hard attention mask from selector input
b, s, h, w = selector.shape
mask = selector.argmax(dim=1).int()
import switched_conv_cuda_naive
output = switched_conv_cuda_naive.forward(input, mask, weight, bias,
stride)
ctx.stride = stride
ctx.breadth = s
ctx.save_for_backward(
*[input, output.detach().clone(), mask, weight, bias])
return output
def create_module(self):
dims = list(self.dims) # convert or copy
if isinstance(self.kernel_size, Iterable):
kernel_sizes = self.kernel_size
else:
kernel_sizes = [
self.kernel_size,
] * len(dims)
if isinstance(self.activation, Iterable):
activations = self.activation
else:
activations = [
self.activation,
] * len(dims)
if isinstance(self.dropout, Iterable):
dropouts = self.dropout
else:
dropouts = [
self.dropout,
] * len(dims)
dims.insert(0, prod(self.pre_dim))
for i in range(len(dims) - 1):
if dims[i + 1] is None:
dims[i + 1] = dims[i]
layers = []
for dim_in, dim_out, kernel_size, activation, dropout in zip(
dims[:-1], dims[1:], kernel_sizes, activations, dropouts):
layers.append(Permute(0, 3, 1, 2)) # (b,l,l,c) -> (b,c,l,l)
padding_t = padding_l = int(np.floor((float(kernel_size) - 1) / 2))
padding_b = padding_r = int(np.ceil((float(kernel_size) - 1) / 2))
layers.append(
ZeroPad2d(padding=(padding_l, padding_r, padding_t,
padding_b)))
layers.append(
Conv2d(in_channels=dim_in,
out_channels=dim_out,
kernel_size=kernel_size))
layers.append(Permute(0, 2, 3, 1)) # (b,c,l,l) -> (b,l,l,c)
if activation is not None:
layers.append(activation)
layers.append(Dropout(dropout))
module = Sequential(*layers)
return module
def __init__(self, input_size: int, output_filters: int, dropout=0.0):
super(UNetUp, self).__init__()
self.model = Sequential(
Upsample(scale_factor=(2, 1)),
ZeroPad2d((0, 0, 1, 0)),
Conv2d(input_size,
output_filters,
kernel_size=(4, 1),
stride=1,
padding=(1, 0),
bias=False),
ReLU(inplace=True),
BatchNorm2d(output_filters, momentum=0.8),
)
if dropout:
self.model.add_module("Dropout", Dropout(dropout))
def __init__(self, pad_size, pre_pad=False, pad_mode='zero'):
"""
Padding which allows to simultaneously pad in a reflection fashion
and map to complex.
Parameters
----------
pad_size : int
size of padding to apply.
pre_pad : boolean
if set to true, then there is no padding, one simply adds the imaginarty part.
"""
self.pre_pad = pre_pad
if pad_mode == 'Reflect':
# print('use reflect pad')
self.padding_module = ReflectionPad2d(pad_size)
else:
# print('use zero pad')
self.padding_module = ZeroPad2d(pad_size)
def init_network(self):
"""Initialize network parameters. This is an actor-critic build on top of a RNN cell. The
actor is a fully connected layer, and the critic consists of two fully connected layers"""
self.rnn = LSTMCell(self.action_space, self.hidden_size)
for p in self.rnn.parameters():
uniform_(p, self.uniform_init[0], self.uniform_init[1])
self.actor = Linear(self.hidden_size, self.action_space)
for p in self.actor.parameters():
uniform_(p, self.uniform_init[0], self.uniform_init[1])
self.middle_critic = Linear(self.hidden_size, self.hidden_size // 2)
for p in self.middle_critic.parameters():
uniform_(p, self.uniform_init[0], self.uniform_init[1])
self.critic = Linear(self.hidden_size // 2, 1)
for p in self.critic.parameters():
uniform_(p, self.uniform_init[0], self.uniform_init[1])
self.encoder = resnet34(**{"num_classes": self.embedding})
self.padding = ZeroPad2d((30, 20, 0, 0))
def data(device="cpu"):
N, C, Hin, Win = 100, 10, 32, 32
padding = [1,2,3,4]
Hout = Hin + padding[2] + padding[3]
Wout = Win + padding[0] + padding[1]
X = randn(N, C, Hin, Win, requires_grad=True, device=device)
module = extend(ZeroPad2d(padding)).to(device=device)
out = module(X)
vout = randn(N, C, Hin, Win, device=device)
vin = randn(N, C, Hout, Wout, device=device)
return {
"X": X,
"module": module,
"output": out,
"vout_ag": vout,
"vout_bp": vout.view(N, -1, 1),
"vin_ag": vin,
"vin_bp": vin.view(N, -1, 1),
}
def __init__(self, base_filters=32):
super(Generator, self).__init__()
self.in_conv = Sequential(
ConvBlock(1, base_filters, stride=1),
ConvBlock(base_filters, base_filters, stride=1))
self.down1 = ConvBlock(base_filters, 2 * base_filters)
self.down2 = ConvBlock(2 * base_filters, 4 * base_filters)
self.down3 = ConvBlock(4 * base_filters, 8 * base_filters)
self.down4 = ConvBlock(8 * base_filters, 8 * base_filters)
self.up1 = Upward(8 * base_filters, 8 * base_filters)
self.up2 = Upward(16 * base_filters, 4 * base_filters)
self.up3 = Upward(8 * base_filters, 2 * base_filters)
self.up4 = Upward(4 * base_filters, base_filters)
self.out_conv = Sequential(
ConvBlock(2 * base_filters, base_filters, stride=1),
ZeroPad2d((1, 2, 1, 2)),
Conv2d(base_filters, 1, kernel_size=4, stride=1), Tanh())
def build_layer_dict(self) -> OrderedDict:
"""Compiles a block-specific dictionary of network layers.
This could be overwritten by derived layers
(e.g. to get a 'BatchNormalizedConvolutionBlock').
:return: Ordered dictionary of torch modules [str, nn.Module]
"""
layer_dict = OrderedDict()
# treat first layer
layer_dict["conv_0"] = self.convolution_nn(in_channels=self.input_channels,
out_channels=self.hidden_channels[0],
kernel_size=self.hidden_kernels[0],
stride=self.hidden_strides[0],
padding=self.hidden_padding[0],
dilation=self.hidden_dilations[0],
padding_mode=self.padding_mode)
layer_dict[f"{self.non_lin.__name__}_0"] = self.non_lin()
# treat remaining layers
for ii in range(1, len(self.hidden_channels)):
padding_to_use = self.hidden_padding[ii]
if isinstance(self.hidden_padding[ii], list) and len(self.hidden_padding[ii]) == 4 and \
self.padding_mode == 'zeros':
layer_dict[f'padding_{ii}'] = ZeroPad2d(self.hidden_padding[ii])
padding_to_use = 0
layer_dict[f"conv_{ii}"] = self.convolution_nn(in_channels=self.hidden_channels[ii - 1],
out_channels=self.hidden_channels[ii],
kernel_size=self.hidden_kernels[ii],
stride=self.hidden_strides[ii],
padding=padding_to_use,
dilation=self.hidden_dilations[ii],
padding_mode=self.padding_mode)
layer_dict[f"{self.non_lin.__name__}_{ii}"] = self.non_lin()
return layer_dict
def generate(self,
phonemes,
len_phonemes,
steps=False,
window=3,
spectrograms=None):
"""Sequentially generate spectrogram from phonemes
If spectrograms are provided, they are used on input instead of self-generated frames (teacher forcing)
If steps are provided with spectrograms, only 'steps' frames will be generated in supervised fashion
Uses layer-level caching for faster inference.
:param phonemes: Padded phoneme indices
:param len_phonemes: Length of each sentence in `phonemes` (list of lengths)
:param steps: How many steps to generate
:param window: Window size for attention masking
:param spectrograms: Padded spectrograms
:return: Generated spectrograms
"""
self.generating(True)
self.train(False)
assert steps or (spectrograms is not None)
steps = steps if steps else spectrograms.shape[1]
with torch.no_grad():
phonemes = torch.as_tensor(phonemes)
keys, values = self.txt_encoder(phonemes)
if hp.positional_encoding:
keys += positional_encoding(keys.shape[-1],
keys.shape[1],
w=hp.w).to(self.device)
pe = positional_encoding(hp.channels, steps,
w=1).to(self.device)
if spectrograms is None:
dec = torch.zeros(len(phonemes),
1,
hp.out_channels,
device=self.device)
else:
input = ZeroPad2d((0, 0, 1, 0))(spectrograms)[:, :-1, :]
weights, decoded = None, None
if window is not None:
shape = (len(phonemes), 1, phonemes.shape[-1])
idx = torch.zeros(len(phonemes), 1,
phonemes.shape[-1]).to(phonemes.device)
att_mask = idx_mask(shape, idx, window)
else:
att_mask = mask(shape=(len(phonemes), 1, keys.shape[1]),
lengths=len_phonemes,
dim=-1).to(self.device)
for i in range(steps):
if spectrograms is None:
queries = self.audio_encoder(dec)
else:
queries = self.audio_encoder(input[:, i:i + 1, :])
if hp.positional_encoding:
queries += pe[i]
att, w = self.attention(queries, keys, values, att_mask)
dec = self.audio_decoder(att + queries)
weights = w if weights is None else torch.cat(
(weights, w), dim=1)
decoded = dec if decoded is None else torch.cat(
(decoded, dec), dim=1)
if window is not None:
idx = torch.argmax(w, dim=-1).unsqueeze(2).float()
att_mask = idx_mask(shape, idx, window)
self.generating(False)
return decoded, weights
in_channels=32,
out_channels=64,
kernel_size=2,
stride=2,
max_alpha=0.), # 1
]
def make_info_layer(idx):
return (INFO_ARGS[idx]['output_size'],
InformationDropoutLayer(Conv2d, **INFO_ARGS[idx]))
SHAPES_LAYERS = [ # (output_shape, layer) tuples. Cross-checked in `.forward`.
((2, 75, 75), None), # Input layer
((2, 76, 76), ZeroPad2d((0, 1, 0, 1))),
make_info_layer(0),
# ((32, 36, 72), ZeroPad2d((0, 1, 0, 1))),
((32, 19, 19), MaxPool2d(kernel_size=2, stride=2)),
((32, 20, 20), ZeroPad2d((0, 1, 0, 1))),
make_info_layer(1),
((64, 5, 5), MaxPool2d(kernel_size=2, stride=2)),
((1600, ), Flatten()),
((2, ), Linear(in_features=1600, out_features=2)),
]
class BentesModel(Module):
"""Implementation of Bentes et.al 2016"""
def __init__(self) -> None:
super(BentesModel, self).__init__()
def __init__(self, in_f, f):
super(ConvStride2NormRelu, self).__init__()
self.zero = ZeroPad2d((1, 0, 1, 0))
self.conv = Conv2d(in_f, f, 3, stride=2, bias=False)
self.bnorm = BatchNorm2d(f, eps=1e-03)
